ABSTRACT
We investigate a novel dataset of more than half a million 15 seconds transcribed audio snippets containing COVID-19 mentions from major US TV stations throughout 2020. Using the Latent Dirichlet Allocation (LDA), an unsupervised machine learning algorithm, we identify seven COVID-19 related topics discussed in US TV news. We find that several topics identified by the LDA predict significant and economically meaningful market reactions in the next day, even after controlling for the general TV tone derived from a field-specific COVID-19 tone dictionary. Our results suggest that COVID-19 related TV content had nonnegligible effects on financial markets during the pandemic.
ABSTRACT
BACKGROUND: Pollen are monitored in Europe by a network of about 400 pollen traps, all operated manually. To date, automated pollen monitoring has only been feasible in areas with limited variability in pollen species. There is a need for rapid reporting of airborne pollen as well as for alleviating the workload of manual operation. We report our experience with a fully automated, image recognition-based pollen monitoring system, BAA500. METHODS: The BAA500 sampled ambient air intermittently with a 3-stage virtual impactor at 60 m3/h in Munich, Germany. Pollen is deposited on a sticky surface that was regularly moved to a microscope equipped with a CCD camera. Images of the pollen were constructed and compared with a library of known samples. A Hirst-type pollen trap was operated simultaneously. RESULTS: Over 480,000 particles sampled with the BAA500 were both manually and automatically identified, of which about 46,000 were pollen. Of the automatically reported pollen, 93.3% were correctly recognized. However, compared with manual identification, 27.8% of the captured pollen were missing in the automatic report, with most reported as unknown pollen. Salix pollen grains were not identified satisfactorily. The daily pollen concentrations reported by a Hirst-type pollen trap and the BAA500 were highly correlated (r = 0.98). CONCLUSIONS: The BAA500 is a functional automated pollen counter. Its software can be upgraded, and so we expected its performance to improve upon training. Automated pollen counting has great potential for workload reduction and rapid online pollen reporting.